Built constraint-aware image generation pipelines using diffusion and inpainting with deterministic slot-filling. Developed and evaluated 6 architectures end-to-end, selecting a final pipeline producing reliable 600 DPI and 300 DPI outputs with compliance checks and export-validity gates.
Designed evaluation harness tracking fidelity metrics (SSIM, PSNR), inference latency, and throughput to guide iteration toward deployment-ready systems. Created and led a team of 11, establishing sprint structure, code-review gates, and reproducible experiment tracking.
Spearheaded NLP pipelines with custom tokenization and Transformer-based sentiment models to convert field feedback into analysis-ready signals. Designed and trained CNN-based image-labeling models achieving 92% validation accuracy for classifying educational materials.
Architected and deployed a WhatsApp chatbot (Twilio API + Python) with Firebase-backed storage, collecting data from over 2000 students. Conducted Hindi/English normalization and topic modeling (LDA + Word2Vec) for impact evaluation. Managed end-to-end M&E workflows for a five-member team and published comprehensive Impact Reports.
Built a mental-health support chatbot processing PHQ-9 and GAD-7 responses through Transformer-based sentiment classifier to generate affective-state vectors, iterated using engagement and response-fidelity checks in IRB-approved pilots.
Developed an agent-based computational model to evaluate policing, screening, and therapeutic interventions for university substance use, validated through sensitivity analyses to support harm-reduction recommendations. Formed and led a team of 5 to execute model development and evaluation.